Multimodal medical AI: A survey of vision–language models, knowledge-enhanced learning, and retrieval-augmented generation in clinical applications

The increasing availability of multimodal data in the medical domain—ranging from medical images and clinical text to Electronic Health Records (EHRs) , and knowledge databases has driven rapid progress in multimodal learning. These models aim to jointly reason over heterogeneous data to improve performance in core clinical tasks such as diagnosis, prognosis, and treatment planning. This survey provides a comprehensive overview and discussion of recent advances in multimodal models tailored for medical applications, covering key tasks such as multimodal Visual Question Answering (VQA) (Med-VQA), medical Image Caption Generation (ICG), disease classification, and segmentation, all aimed at supporting clinical decision-making. We further examine emerging paradigms such as Retrieval-Augmented Generation (RAG) and Knowledge Graphs (KG)-enhanced models, which incorporate external medical knowledge to improve reasoning, robustness, and interpretability. In the survey, we identify the challenges and opportunities of multimodal integration in medicine, detail representative model architectures and learning strategies, and analyse their strengths and limitations across tasks and models. We also outline promising future research directions, including knowledge-aware reasoning, scalable foundation models, and clinically grounded evaluation protocols. We believe this survey will serve as a valuable reference for researchers and practitioners developing next-generation multimodal medical Artificial Intelligence (AI) systems.